Fair Allocation of Bandwidth at Edge Servers for Concurrent Federated Learning Processes

Md Anwar Hossen, Fatema Siddika, and Wensheng Zhang

10th Int. Conf. on Fog and Mobile Edge Computing (FMEC) 2025

TL;DR. When several federated learning jobs share the same edge servers, bandwidth becomes contested. This work models that contention as a game and derives distributed heuristics that approximate a fair, efficient allocation.

Motivation

Edge servers increasingly host concurrent federated learning processes that all compete for limited uplink/downlink bandwidth. Without a principled allocation, some jobs starve while others over-consume, hurting overall efficiency and fairness.

Approach

We formulate bandwidth allocation across concurrent federated learning processes as a game-theoretic problem and design distributed heuristics that approximate a Nash equilibrium. The result is an allocation that improves bandwidth utilization at the edge while remaining fair across competing learning processes and that runs without central coordination.

Game-theoretic formulation of fair bandwidth sharing at the edge.
Distributed heuristics approximating an equilibrium allocation.
Improved, fairer bandwidth utilization for concurrent FL jobs.

Origin

This project grew out of my M.S. research at Iowa State University and was recognized with a Best Research Poster award in Computer Science.